Art Market Intelligence. AI + Human Judgment. Real Impact.
Live: bjbeyond.it
| Framework | Next.js 16 (App Router), static export |
| Language | TypeScript, strict |
| Styling | Tailwind CSS v4 (CSS-first @theme tokens) |
| Motion | CSS transitions on one shared IntersectionObserver · GSAP + ScrollTrigger for desktop scroll choreography, loaded on demand |
| Deploy | GitHub Pages, from .github/workflows/deploy.yml |
npm installnpm run devThen open http://localhost:3000.
| Script | Does |
|---|---|
npm run dev |
Dev server with HMR |
npm run build |
Static export into out/ |
npm run typecheck |
tsc --noEmit |
npm run media |
Regenerates public/media/ from media-src/ |
There is no
lintscript, on purpose. There used to be one callingnext lint, which Next 16 removed, against an ESLint configuration that did not exist — so everyeslint-disablecomment in the tree was suppressing a rule nothing was running. Its replacement,eslint-config-next's flat config, cannot run here either: it loadstypescript-eslint, which refuses to start against the TypeScript 7 this project builds with (typescript-eslint#10940). Rather than leave 300 packages installed that error on launch, the script is gone until that lands.npm run typecheckis the gate in the meantime, and withnoUnusedLocalsandnoUnusedParametersit is a strict one.
On
/phoenixand/frequency: these are directories inpublic/containingindex.html. Static hosts resolve a directory to its index automatically, so both work in production.next devdoes not, sonext.config.mjscarries dev-only rewrites to match. The build prints a warning that rewrites are ignored when exporting — that is expected and correct; they exist purely for development parity.
app/ layout, homepage, the inner pages, global tokens
art/ books/ method/ services/ the pages the menu leads to
labs/ contact/
privacy-policy/ cookie-policy/ legal, own shell
sitemap.ts robots.ts generated from lib/routes.ts
schema.json/ generated from lib/schema.ts
components/
atmosphere/ procedural photography substitute (see below)
chrome/ nav, intro, cursor, scroll sync, logo, page shell,
consent banner, analytics
primitives/ reveal, magnetic, arrow link, rail
sections/ hero, about, method, authentia, in-motion,
labs, work, intelligence, contact, footer
lib/
content.ts ALL copy — single source of truth
routes.ts ALL urls — single source of truth
schema.ts JSON-LD, derived from content.ts
analytics.ts measurement ids, and the consent rule over them
media.ts image manifest — the only place photography enters
motion.ts easing + duration tokens
reveal.ts the shared IntersectionObserver behind every entrance
gsap.ts the desktop gate, and the dynamic import behind it
gsap-runtime.ts the library itself — never imported statically
public/ legacy apps and assets, copied verbatim into the build
_legacy/ the pre-rebuild index.html, kept for reference only (not built)
The site was one document with an anchor menu. It is now a homepage plus a set of real URLs, because an anchor cannot be indexed, cannot carry its own title or link preview, and cannot be the destination of an ad.
lib/routes.ts is the single source of truth for URLs. The
sitemap, the navigation, ArrowLink's decision between next/link and <a>,
and every page's metadata all read it. Adding a page means adding an entry there
and creating app/<name>/page.tsx with:
export const metadata: Metadata = metadataFor('/name/');That one call produces the title, description, canonical, Open Graph card and Twitter card together — they have to agree, and doing them by hand per page is a fresh chance to get one wrong per page.
Two rules that are easy to break:
- Trailing slashes are part of the path.
trailingSlash: truemeans the served URL is/method/. A manifest entry or canonical without the slash names a URL that redirects. draft: truekeeps a page out of the menu and out of the sitemap, and marks itnoindex./books/carries it today, because its three entries are placeholders the owner asked for. Announcing a URL that 404s, or letting Google index invented book titles under a real author's name, are the two things this flag prevents.
The homepage keeps the full scroll narrative — it is the brand experience and
the landing for cold traffic. The inner pages are flatter on purpose: someone
arriving from an ad wants the answer, not the entrance, which is also why
PageShell mounts IntroProvider with curtain={false}.
Links between pages are plain <a>, never next/link. That is a
measurement decision, not a routing one, and it is load-bearing — see below.
Every page carries one <script type="application/ld+json"> describing an
Organization and a WebSite, built by lib/schema.ts from
SITE, SOCIAL and CONTACT. This is the vocabulary Google reads to treat
BJ Beyond as an entity rather than a phrase, and sameAs is what attaches
the X, TikTok, Threads, Reddit and Substack accounts to that same entity.
There used to be a public/schema.json doing none of this. Nothing referenced
it — no script tag, no link — so no crawler could find it, and it had drifted
into naming a social account the owner had replaced, an image that does not
exist, and a contact property that is not in the schema.org vocabulary at all
(the spelling is contactPoint; an unknown term is ignored, not reported).
It is now derived, not written, which is the point: structured data is a
second copy of facts the site already states, and a second copy that is
maintained by hand is the kind that goes quietly wrong. Change a social link in
lib/content.ts and the JSON-LD changes with it.
The /schema.json URL still resolves — app/schema.json/route.ts
serialises the same object — so any old link keeps working without there being
two definitions of these facts.
Not included, deliberately: the owner's legal name as founder. It appears
in the privacy policy because a data controller must be named there; publishing
it as machine-readable metadata on every page is a different act, and the
owner's call. It is one property when they want it.
Nothing measurable loads before consent. Not GA4, not the Meta pixel — no request is made to Google or Meta until the visitor accepts.
GA4 previously ran on every visit with no banner and nothing to refuse it. That was a gap on its own; adding an advertising pixel to it would have made the cookie policy's own statement — that the site uses no advertising profiling cookies — untrue.
| Piece | Where |
|---|---|
| Measurement IDs, consent signals | lib/analytics.ts |
| Banner, stored choice, withdrawal | components/chrome/Consent.tsx |
| The tags, and per-route page views | components/chrome/Analytics.tsx |
| Consent Mode v2 defaults (denied) | inline in <head>, app/layout.tsx |
-
The Meta pixel is live. Setting
META_PIXEL_IDalso publishes the cookie policy's marketing rows and its extra-EU transfer notice, which are derived from that same constant inlib/legal.ts. The document cannot describe a pixel the site does not run, or stay silent about one it does. Setting it back tonullwithdraws both together. -
No page-view code exists, on purpose. Every internal link is a plain
<a>, so each page is a fresh document load and both tags do their own counting — which is what they are built for.This replaced three failed attempts at driving them by hand under
next/link, each of which failed silently: manual sending doubled every navigation (GA4's Enhanced Measurement was already tracking History API changes),send_page_view: falseremoved the landing page instead of the duplicate, and leaving GA to Enhanced Measurement alone lost a third of the navigations. Measured by interceptingsendBeacon/fetch/XHR— resource timing undercounts, because GA4 batches events into one request.If client-side navigation is ever wanted back, the manual sending has to be rebuilt and re-measured on the live domain. Not on localhost: the Meta pixel reports the origin rather than the page path there, so it cannot be verified. The full reasoning is in
Analytics.tsx. -
Expect lower numbers than before. Visitors who ignore the banner are now invisible. The old figures were not more accurate — they were collected without asking.
Source art lives in media-src/ as bitmaps wrapped in SVG envelopes:
| File | Contains |
|---|---|
background.svg |
A four-panel contact sheet, 7680×4352 |
avatar.svg |
The cut-out portrait, RGB + alpha mask layers |
logo.svg |
The white BJ monogram with alpha |
npm run media turns these into the responsive WebP variants the site serves —
splitting the contact sheet into four scenes, cropping away its burned-in
reference labels, and emitting each at several widths. All the imagery on the
site totals ~790 KB, down from 6.6 MB of source.
Two things in that script are deliberate and easy to get wrong:
- The background is read by extracting the embedded JPEG, not by rasterising the SVG. Rendering the wrapper applies its own viewBox placement, which reframes the sheet so the panel grid no longer lines up.
- The portrait is the opposite: it must be rasterised, because the envelope holds a colour layer and a separate greyscale alpha mask, and only the renderer composites them into the cut-out figure.
Slots still without a photograph — the three Labs plates — render a
procedural atmosphere instead: layered ridge silhouettes, drifting fog, a
warm source and grain, generated deterministically from the slot id. To give one
a real image, add its variants to public/media/ and set src + widths on
that slot in lib/media.ts. Nothing else needs to change.
These URLs predate the rebuild and still resolve exactly as before:
/phoenix— Phoenix Simulator/frequency— Frequency Studio/pages/privacy-policy.html,/pages/cookie-policy.html/docs/*,/CNAME/schema.json— same URL, no longer a static file; see Structured data
/assets/* was on this list and is gone. It held four files — a 14.7MB MP4 and
three images — that nothing in the site referenced: not the sections, not the
standalone apps, not a stylesheet. They were 85% of every deploy. Git history
still has them if one is ever wanted back.
They are served verbatim from public/ and are not managed by the router.
Link to them with a plain <a>, never next/link — they are not App Router
routes and next/link would try to resolve them client-side.
Phoenix and Frequency are deliberately untouched. Their markup, styling and behaviour are byte-identical to the pre-rebuild files; they keep their own design. Do not restyle them to match the main site unless explicitly asked.
The privacy and cookie policies were rebuilt, and now live as real routes:
/privacy-policy—app/privacy-policy/page.tsx/cookie-policy—app/cookie-policy/page.tsx
Both render through components/legal/LegalPage.tsx
from structured content in lib/legal.ts. The wording is
transcribed verbatim from the original documents — it is a legal statement about
data handling, so only the presentation changed.
The old /pages/*.html files are still served untouched, so any link already
published elsewhere keeps working. The site's own footer points at the new
routes.
GitHub Pages, and only GitHub Pages. .github/workflows/deploy.yml
installs, builds, and uploads out/.
Two things that have each broken the live site once:
- Pages' source must stay on GitHub Actions, not "Deploy from a branch".
On a branch source, Jekyll rebuilds from the repository root, finds no
index.html, and renders the README as the homepage. public/.nojekyllis required. Without it Jekyll strips_next/, which removes every stylesheet and script.
vercel.json and wrangler.jsonc used to sit alongside this for platforms
nobody deploys to — the Cloudflare one registered under a different project's
name. They were removed rather than maintained; both were three lines, and git
still has them if a second target is ever wanted.
- Art Market Pulse data is illustrative, not a live market feed. The interface says so on screen. Do not present it as real market data.
- API keys are not hardcoded. The Phoenix simulator and worker read them
from user-supplied
<input type="password">fields at runtime. (An earlier version of this README warned otherwise; that warning was out of date.) - The BJ monogram in
components/chrome/Logo.tsxis a geometric reconstruction, drawn as SVG because the bitmap in/assetsis dark-on-dark with no alpha and cannot be shown white on a dark background. Replace its two paths with the official vector when available.
One step beyond AI.